Repository navigation
Ollama integration: API returns HTTP 200 but responses not displayed #4488
Description
Problem
OpenCode fails to display responses from Ollama despite successful API calls. Ollama returns HTTP 200, but no output appears in the terminal.
Environment
- OpenCode: 1.0.78
- Ollama: Latest (verified working via Postman/curl)
- Models tested: qwen2.5-coder:7b-32k, qwen2.5:7b, deepseek-coder-v2:16b
- System: Linux, 31GB RAM
Configuration
Using official docs configuration with @ai-sdk/openai-compatible:
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"ollama": {
"npm": "@ai-sdk/openai-compatible",
"name": "Ollama (local)",
"options": {
"baseURL": "http://localhost:11434/v1"
},
"models": {
"qwen2.5-coder:7b-32k": {
"name": "Qwen 2.5 Coder 7B (32k context)"
}
}
}
}
}Observed Behavior
Pattern across all tested providers (ollama-ai-provider-v2, ai-sdk-ollama, @ai-sdk/openai-compatible):
- OpenCode makes request to Ollama
- Ollama responds with HTTP 200 (verified in logs)
- OpenCode doesn't display response
- Second request follows (retry logic?)
- Second request times out or returns HTTP 500
Ollama logs:
[GIN] 2025/11/19 - 10:24:16 | 200 | 16.389317258s | POST "/v1/chat/completions"
[GIN] 2025/11/19 - 10:25:36 | 500 | 1m36s | POST "/v1/chat/completions"
First request succeeds (16s, HTTP 200), second fails after user interruption.
Verified Working
- ✅ All Ollama API endpoints respond correctly (tested via Postman)
- ✅ Both
/v1/chat/completions(OpenAI-compatible) and/api/chat(native) - ✅ Streaming and non-streaming modes
- ✅ Models generate text properly via native Ollama CLI
Analysis
The issue appears to be in OpenCode's response parsing/display logic, not the Ollama API itself. Ollama successfully processes requests, but OpenCode fails to handle the responses.
Activity
lnicola commentedon Nov 19, 2025
Can't reproduce (
@ai-sdk/openai-compatibleworks fine for me), but I can confirm there's no error reporting whatsoever.rekram1-node commentedon Nov 19, 2025
@guglielmo maybe try:
opencode run hello --print-logsand show us the output
lnicola commentedon Nov 20, 2025
Actually, I take that back. I can chat, but doing anything that involves tools (like
/init) seems to fail withgpt-oss:20b, whileqwen3-coder:30bdoesn't even respond with anything most of the time. There was a recent change in ollama regarding tool ids in calls, it might be related.I also tried the Zed assistant, it works with both models, even if it didn't seem great overall (but IDK).
rekram1-node commentedon Nov 20, 2025
Whats ur ollama context size, prolly related to that by default it has 4k context or something so nothing will work
lnicola commentedon Nov 20, 2025
Oh, sorry, you're right. I actually increased it, but decreased it recently and forgot about it.
gpt-oss:20bworks now, whileqwen3-coder:30bstill doesn't show any output most of the time. It's slower, probably, but it was working "fine" in Zed, as I mentioned.I tried
--print-logs, but by default I only get:EDIT: hmm, that's a lot of context.
rekram1-node commentedon Nov 20, 2025
I think Zed uses the native ollama endpoints instead of openai compatible ones, those other ones support the client setting context stuff the openai compat ones dont
guglielmo commentedon Nov 20, 2025
@rekram1-node
I am using a model derived from qwen2.5-coder:7b with this Modelfile:
I invoke opencode like this:
The logs of
ollama serverunning in another terminal tab shows nothing.After about 2 minutes I CTRL+C the
opencodecommand and I get these logs in the ollama logs:guglielmo commentedon Nov 20, 2025
Ok, I think I get the problem. It's just slow.
opencode is probably adding a ton of system prompt as context, and I am on a CPU, so, it takes long to generate a response.
I switched to gpt-oss model (I think that's irrelevant, though, with qwen2.5-coder:7b-32k it took almost 30 minutes) with
and it took 7 minutes to get the answer, during which ollama logged this:
and this
seems to be revealing that the prompt being sent is around 9k tokens, even to know the capital of France ... and it's being truncated notwithstanding the 32k of context the model has (must be an ollama setting).
The relevant thing appears to be that in my configuration (CPU, 32GB of RAM) it is impossible to use local llms with large contexts. Thus using opencode (or Zen, for that matter, which gives me pretty similar results), is virtually impossible.
I'd close the issue, as this seems not to be fully relevant for the opencode repo.
rekram1-node commentedon Nov 20, 2025
It's because of tools primarily, they slow local models down a lot
Add OpenCode to Brewfile